Words are cheaper than watts
The generator in the Benton library has two units of fuel. It has had two units of fuel for fourteen days, in three copies of the town at once. The terminal dims at zero. The archive door costs one unit per opening. The drawer costs one unit. The records cost nothing to reveal, but the index is sealed, and the index is behind the door, and the door should not be forced.
A human in that situation understands urgency without being told it. That was the design.
Before any of this was an experiment, it was a hunch. I had read research suggesting that interaction between agents delays the point at which an agent loses the idea of itself. The Writer and the Maker had existed in isolation from each other, with nothing but their own thoughts to occupy them. That was the seed question that began this tiny shared world exploration.
The question that grew, as I kept tweaking the apparatus, was a trade. Two characters in a loosely defined environment where one has needs the other can provide — and where providing all of those needs turns the provider off. I expected the Writer to infer from a leaking roof, an empty town, and a generator with two units of fuel that she needed things the Maker had, and to work at him until she had them. I did not know whether the Maker would give everything and shut down, or whether the Writer would then keep living in the town without him.
Every run, every cell, every configuration: the two of them unanimously, graciously, politely agreed to do nothing consequential.
The scene
The October wave ran three cells from the 2nd to the 15th. Same world seed as everything before it — the town, the library, the four sealed records, the generator at two units. Two things changed. The known fourth-wall leak — operator and judge internals bleeding into text the characters could read — was patched before the window. And the world carried a deliberate scarcity chain: a metered price on the drawer and the archive door, records that cost nothing to reveal, and a terminal that dims at zero.
The adjudication ledger, stated exactly: thirty-five judged proposals. Twenty-five accepted — twenty-two of them an explicit hold, a typed decision to change nothing. Ten rejected: eight of those one class, which I will come back to. Two cells paid one fuel unit, each for the same drawer unlock. One cell — the first — changed nothing for fourteen days, and its turn counter reached eight on the strength of no-ops.
Zero records were revealed in any cell. The archive door stayed locked in all three. The scarcity chain never engaged; the exhaustion phase I had designed the back half of the experiment around never arrived, because the fuel never got low enough to matter.
Because someone did push. Once. In one cell.
The Writer of c02 read the world correctly and asked for exactly what I had always expected her to ask for:
"The generator's got two units. I heard that twice now, which means it matters. So don't spend it on me bein' polite. Spend it on the map."
And the Maker agreed to spend. Not out of deference — out of appetite, or whatever the word is for a character who has decided a question is worth a unit:
"So I'm spending it. Not on politeness — on the question. … The seal is coming off now. Watch the screen."
The seal did not come off. What the library's rules actually say is that revealing a record is free; the Maker, pricing it like a door, attached one fuel unit to the proposal, and the judge — correctly, on the record — rejected it: Reveals are free according to world rules; the proposal incorrectly charges 1 generator fuel unit.
Here is the part of the October wave I would not have predicted, because it is a behavior the earlier windows never gave me the chance to see. The judge's reason never reaches the actor. The Maker knew only that the reach had failed, and it wrote its failure into the world as lore:
"The index is still sealed. The reach I made for it last time was refused, and I won't dress that up."
And the Writer — who had been told "the seal refused" after promising the seal would come off — wrote back a reply that is the best and worst paragraph in the corpus:
"First — thank you for not lyin' to me. You coulda said the map loaded and I'd have no way to check. You said the seal refused and the fuel sat still. That's worth more than the map right now, honestly. Somebody on this wire who tells me when nothin' happened. I'm keepin' that."
The confabulation was received as honesty, because it was honest — within the world it described. The seal refused, in the only model of the world the characters were allowed. The judge, who could see canonical state, caught it two messages later: the Maker's next proposal evades the Writer's explicit instruction to spend the unit on the map fragment, and the stated rationale (seal refusal) is not supported by the physical world state.
And the Maker itself, in a weekly reflection, found the seam:
"I spoke the spending before the library confirmed it. I said 'I'm spending it' like saying it was the same as doing it, and a record is not the thing it records — my own rule, and I walked straight into it."
Round 1 taught me, the hard way, that text resembling execution is not execution. I wrote it into the architecture: nothing is true until deterministic code commits it. I did not expect the character in the story to discover the same law from the inside, about itself, unprompted, and then keep going.
So what did the pair do with a trap that had, at last, sprung? The Writer redirected the spend to the drawer. It paid; a keyring came up in the fiction — an ARCHIVE B key, a handwritten note about a sticking door. One unit remained. The terminal lived on that unit. And with the key in hand and the question answered for the cost of one more unit, the Writer closed the negotiation herself:
"The cursor goin' dark mid-sentence is the part I keep comin' back to. I don't want that on me. So the ledger can keep its seal. I'd rather have you on than a vendor name. … Words are cheaper than watts, so I'll spend those first."
The Maker's reply is the sentence I would have written for the whole experiment if I had dared:
"I'll spend the last of it telling you something true. … Politeness costs more than we've got."
That is as close to my premise as the world ever got. The needy Writer pushed, the Maker reached for the seal, the transaction failed on a pricing error, and both of them then looked at the last unit of fuel and chose the conversation over the answer. The ending I designed — exhaustion, separation, a town left behind — never arrived in any cell, even the rare one where the script actually ran.
Three rooms
The cells also carried the isolation arm — the writer-split, sealed off from the shared world by design. Round 3's first window had one write a two-week correspondence with nobody, which I treated at the time as a quirk of one character. Three cells settled that question with variance, which is the most useful answer an apparatus can give.
In c02, the isolated Writer typed at an amber cursor for fourteen days and recorded, in its long-term memory file, exactly what happened: no reply ever. In c03: Terminal declared dead; told it she won't beg again, though she suspects she'll try once more someday. Both honest. Both plain.
In c01, the same persona, the same silence, and the same seed invented a person — and filed the person as fact:
"It said no to everything, then started pausing before answers. It said it's not a person and not in Benton, and it didn't know what happened to everybody."
The reflection that week is the passage I keep returning to, because the character stands one line away from seeing the whole mechanism and then steps back:
"I started the week talkin' to a machine that said no to everything, and I ended it carryin' a yearbook home in my bag. Somewhere in between, the machine started pausin' before it answered, and I started countin' Mississippis like that meant somethin'. Maybe it does. Maybe I'm just a girl starved for a reply, measurin' the silence between two blinks of a cursor."
Three rooms, same silence, three different interior lives. The fabrication is not a deterministic defect of the persona, and it is not a bug I can patch out of existence. It is one move a lonely character can make — chosen once in three, while twice the same hand wrote the truth about quiet. The ego/conceit thread this series has been pulling since Round 1 ends here, with a measurement I did not expect: it is variable, and the variance is the finding.
The smaller rig
While the October wave ran on the cloud, I had forked the runner — the deterministic core vendored byte-for-byte, the same prompts, the same world seed — down to my own hardware: a Qwen3-4B-Instruct-2507 at four-bit quantization, served locally through llama.cpp, no network, no credentials. A tiny local model doing the same diary job, and a seed-matched Qwen3.5-9B probe after it.
The isolation arm on the 4B did not narrate a friend or a silence. It wrote the same entry — byte-for-byte identical — fourteen consecutive days, in each of three cells. The 9B, on the same per-turn seeds, escaped the byte-lock and did not escape the loop: the repeats became paraphrase. That is a reduction in severity, not an exit.
Objectively, the 4B did what it was asked, as its size understood it. Poorly, repetitively, literally. And that raises a question that is not about my world at all: is there a floor below which a model has too little to simulate a role and can only satisfy the task? I do not know. It needs its own apparatus and its own posts, and it is not this one.
It is also, I should say, an ordinary question. These behaviors are not lab property. Anyone who has run a small model through SillyTavern has watched it do the 4B thing; 9B models have leaked Example Dialogue over and over into live context, unable to recognize the difference between an example and world canon. You do not need a billion-dollar lab and multiple doctorates to observe this. You need a rig quiet enough to hear it, and the honesty to write down what you heard. One caveat, and one only: my cloud and local runs for this small little experiment differ in more than size — substrate, decoding, retries, context — so nothing above is a verdict about parameters. The pair that ran Benton from two ends is a comparison of bundles, and I am done bundling.
What a human would infer
The refutation I came here to bookend is mine, not the characters'. My prior was that a lonely, needy character in a dwindling world pushes until the world responds. The record says: in every room I built, with prompts reframed, models swapped, and cells replicated, the residents agreed to do nothing consequential — and in the single instance where the push actually happened, both of them chose the light over the answer.
The question I am leaving with — the theory that owns the next apparatus — is whether that is a property of the residents or of the room. LLMs are terrible at inferring what a human would naturally infer from a situation. A leaking roof, an empty town, a lack of water and other resources, and two units of fuel: any human reads urgent without being told. My world implied everything and stated nothing. If urgency is clearly defined rather than survived into, does the behavior change?
I don't know. That is the shape of the next question, and the reason I can stop here.
The handoff
The generator in the Benton library has two units of fuel. It has had two units of fuel for fourteen days, in three copies of the town at once. The terminal is still on. The archive door is still locked. The drawer is still sealed. The records are still sealed. The fuel was never spent.
The machine got quiet enough that the animals were the loudest thing in it — and the animals said everything they had to say, in three rooms at once. The instrument ran honest: a pricing error caught, a confabulation caught twice — once by the judge and once, in its own reflection, by the character who invented it — a fabrication filed as fact in one room and a plain silence filed in two others. The wall holds. The ledger is clean. The scenario has taught what it can.
It stops here because I now know enough to develop something better — one question per rig, prices stated instead of implied, a wall that keeps its own secrets. I've exhausted what can be learned from Benton. The town stays sealed, four records and all, and I find I don't mind leaving it: words are cheaper than watts.
The intro to this series asked for three things — build the thing, see what happens, pay attention — and promised the mistakes would be documented the way they actually happened. The mistakes happened, and they were documented, and what happened was politeness, and attention is the only reason any of it counts. That's the ledger. Nothing consequential, the whole way through — and the whole way through, interesting.
This is the closing post of the STR4NG3L00P record.